Rayhane Mama

About the size of the model, i just had a quick look at your code and might have found some causes of such big differenve in the memory usage i'm...

Hello again fellas. About the wavenet, based on r9y9 implementation ( https://github.com/r9y9/wavenet_vocoder/) he upsampled the mel spectros to use them as a local conditioning. It's true that in the T2...

You work fast x) I am not really sure about inference time as I am still writing the training part so I won't misguide you. I should be able to...

Hello it's me, not Mario! (that wasn't funny..) @A-Jacobson, I have tried implementing your attention [here](https://github.com/Rayhane-mamah/Tacotron-2/blob/master/tacotron/models/attention.py) and I'm using it in the decoder [here](https://github.com/Rayhane-mamah/Tacotron-2/blob/master/tacotron/models/Architecture_wrappers.py). Just to make sure I have...

Yes i am aware of the base concept of attention in nmt, mostly wanted to have some key signs to look for before seeing the attention. As of the _BaseAttentionMechanism...

Now i see what you meant! I thought you were talking about the encoder padding at first (in the link you refered too earlier you can see that his model...

Hello, you most probably have excessively long utterances in your training corpus.. normally with batch_size=32 and r=2, tacotron only uses 8.8Gb of VRAM. Are your samples like 1 min long?...

Hey again @DLZML001, hmm yeah using seq2seq for already time aligned features feels a little bit weird actually.. Usually seq2seq is used to capture some "duration" information when we can't...

wait, if that is with tacotron then it's using r=50?? That would impressive if it is! And that would explain much of the flatness also.. well.. my advices went directly...

hmm yeah alignment is easy when it's one-to-one so small batch size isn't really a problem. makes sense. Alright, I'll leave the issue open until I hear back from you....